Improving synthetic CT accuracy by combining the benefits of multiple normalized preprocesses

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate synthetic computed tomography (sCT) generation is crucial for adaptive radiotherapy.
• Current sCT methods face challenges in accurately representing diverse tissue types.

Purpose of the Study:

• To evaluate the impact of various normalization preprocessing techniques on deep learning-based sCT accuracy.
• To develop a combined approach to enhance the accuracy of all tissues in sCT images.

Main Methods:

• Utilized Cycle-Consistent Adversarial Network (CycleGAN) to generate sCT from megavolt cone-beam CT (MVCBCT).
• Applied seven normalization methods (two linear, five nonlinear) to training data.
• Developed sCT_Blur by merging and smoothing sCT images from different normalizations.

Main Results:

• Different normalization methods improved sCT accuracy for specific tissues.
• The proposed sCT_Blur method demonstrated superior accuracy across all tissues compared to single normalization methods.
• sCT_Blur achieved a structural similarity of 0.906 ± 0.019 to CT and reduced mean absolute errors for key organs.

Conclusions:

• Combining normalization preprocessing methods effectively enhances sCT accuracy for all tissues.
• The sCT_Blur approach shows significant promise for improving dose calculations in adaptive radiotherapy using CBCT data.